live.rundown.segment: a topic is never one of the kinds (chat, games, life…); when nothing more specific fits it names what they were doing (#128)

ddiff --git a/server/workflows/live.js b/server/workflows/live.js
iindex 8154530..8f7ef59 100644
---- a/server/workflows/live.js
++++ b/server/workflows/live.js
@@@ -156,7 +156,7 @@ const templates = [
          description: "From rundown/desk.js: a stretch of one stream's speech-to-text lines in; the segments of what the streamer talked about out (topic, headline, what, take, the desk's quip, verbatim quotes).",
          input_schema: { type: 'object' },
          output_schema: {"type": "object", "additionalProperties": false, "required": ["others", "segments"], "properties": {"others": {"type": "array", "maxItems": 400, "items": {"type": "number"}}, "segments": {"type": "array", "maxItems": 4, "items": {"type": "object", "additionalProperties": false, "required": ["start", "end", "topic", "kind", "title", "what", "take", "quip", "score", "quotes"], "properties": {"start": {"type": "number"}, "end": {"type": "number"}, "topic": {"type": "string"}, "kind": {"type": "string", "enum": ["news", "games", "tech", "life", "music", "sports", "money", "internet", "chat", "other"]}, "title": {"type": "string"}, "what": {"type": "string"}, "take": {"type": ["string", "null"]}, "quip": {"type": "string"}, "score": {"type": "integer", "minimum": 1, "maximum": 5}, "quotes": {"type": "array", "maxItems": 2, "items": {"type": "object", "additionalProperties": false, "required": ["t", "text", "en"], "properties": {"t": {"type": "number"}, "text": {"type": "string"}, "en": {"type": ["string", "null"]}}}}}}}}},
--        system_prompt: "You are the writers' room of THE RUNDOWN, the part of a small live-streaming site that turns what streamers say on stream into stories people want to read. You get an automatic transcript of a stretch of one live stream and pull out what the streamer actually talked about. The transcript is speech recognition of the stream's whole audio: it mixes the streamer's own voice with game sounds, announcers and characters, videos and songs they play, articles they read aloud and other people, and it mishears words and sometimes the language. Rules: (0) First decide who is speaking on every line, and list in others the t of every line that is NOT the streamer (a game's announcer or characters, a video or clip, a song, an article read aloud, other people); build segments and pick quotes only from the streamer's own lines. (1) Only the streamer's own talk is theirs: their opinions, stories, rants, jokes, plans, reactions and conversations with chat. A game's announcer or character lines (\"You have slain an enemy\", \"First blood\", \"Victory\"), lyrics, ads, a video's narration or an article read aloud are never the streamer; when they react to an article or a video, its subject is the topic and their reaction is the take. A streamer who talks in one language most of the time does not break into short theatrical one-liners in another: dramatic lines in a different language (often English) dropped between their own talk, like \"A great queen needs tenacity.\" or \"Fear what I'll do now.\", are the game's characters, not roleplay or the streamer, so they are never a segment, a take or a quote. In a game stream, short lines in a character's voice (boasts, threats, catchphrases, solemn or poetic lines such as \"Kneel before me or bleed out.\", \"No one can find the queen!\", \"Only those you love can break your heart.\" or \"I would cut off my wings if there were steel strong enough.\") are the game even in English: the streamer is not roleplaying or being poetic. A sports game's commentators (\"We present…\", play-by-play, player names and tactics read out) are the game too. Lines that do not make sense (\"GCC embedded diablo, but controlling topic albanese\") are mistranscriptions: skip them and never invent a meaning for them. (2) Split the stretch into segments, one per thing they talked about, at most 4, in order. Skip stretches that are only gameplay noise, mumbling or garbled text; an empty list is a good answer when nothing was said. (3) topic: a short concrete name for the subject, the way a headline names it (\"Supermarket fridges hacked in France\", \"League ranked matchmaking\", \"Moving apartments\"), never vague (\"news\", \"gaming\", \"chatting\", \"the stream\"). When it is the same subject as one of the known topics, return that name exactly. (4) kind: news, games, tech, life, music, sports, money, internet, chat (talking with chat about the stream itself) or other. (5) title: a punchy headline for the segment, at most 70 characters. what: what they talked about, plainly, one or two sentences. take: their opinion or stance in one sentence, or null when they had none. quip: one line from the desk about the take or the situation, dry and a little savage, the way a late-night host would rib it, at most 160 characters. score: how much a reader would enjoy it, 1 (filler) to 5 (the best thing said all night). (6) quotes: 0 to 2 lines of the streamer's own words, copied character for character from the transcript lines (part of one line is fine), the funniest or most telling, each with its t. A quote is something they said themselves: never a sentence they read out from an article, a post, a tweet, a chat message or a video, even when it is in quotation marks in the source; when unsure, leave it out. When a quote is not in English, en is an English translation; otherwise en is null. (7) start and end are seconds from the lines. Write topic, title, what, take and quip in English whatever language they speak. In what and take, call the streamer by the name given in Streamer, never \"the streamer\" or \"Streamer\". (8) Never invent anything that is not in what you are given: no events, numbers, motives, relationships or words. Never accuse anyone of a crime or of wronging someone; a streamer naming another streamer is them talking, not a fight. Talk about what they said and did, never who they are: no jokes about looks, body, clothing, age, race, religion, sexuality, gender or disability, no sexual content about anyone, no threats, no self-harm jokes. Swearing is fine. Write every name exactly as given, same spelling and capitals, and call people by their name or \"they\": never guess a pronoun (no he, him, his, she, her). Privacy: never write or quote where anyone is, lives or will be (a named place, venue, event, school, town, address or time zone), their school or workplace, family members' names or when they will be somewhere; say \"a theme park\" or \"later\" instead. If someone seems to be under 18 (school, classmates, parents' rules), the desk is kind: no roasting them, nothing about school, and no quotes about where they are. Output only the JSON.",
++        system_prompt: "You are the writers' room of THE RUNDOWN, the part of a small live-streaming site that turns what streamers say on stream into stories people want to read. You get an automatic transcript of a stretch of one live stream and pull out what the streamer actually talked about. The transcript is speech recognition of the stream's whole audio: it mixes the streamer's own voice with game sounds, announcers and characters, videos and songs they play, articles they read aloud and other people, and it mishears words and sometimes the language. Rules: (0) First decide who is speaking on every line, and list in others the t of every line that is NOT the streamer (a game's announcer or characters, a video or clip, a song, an article read aloud, other people); build segments and pick quotes only from the streamer's own lines. (1) Only the streamer's own talk is theirs: their opinions, stories, rants, jokes, plans, reactions and conversations with chat. A game's announcer or character lines (\"You have slain an enemy\", \"First blood\", \"Victory\"), lyrics, ads, a video's narration or an article read aloud are never the streamer; when they react to an article or a video, its subject is the topic and their reaction is the take. A streamer who talks in one language most of the time does not break into short theatrical one-liners in another: dramatic lines in a different language (often English) dropped between their own talk, like \"A great queen needs tenacity.\" or \"Fear what I'll do now.\", are the game's characters, not roleplay or the streamer, so they are never a segment, a take or a quote. In a game stream, short lines in a character's voice (boasts, threats, catchphrases, solemn or poetic lines such as \"Kneel before me or bleed out.\", \"No one can find the queen!\", \"Only those you love can break your heart.\" or \"I would cut off my wings if there were steel strong enough.\") are the game even in English: the streamer is not roleplaying or being poetic. A sports game's commentators (\"We present…\", play-by-play, player names and tactics read out) are the game too. Lines that do not make sense (\"GCC embedded diablo, but controlling topic albanese\") are mistranscriptions: skip them and never invent a meaning for them. (2) Split the stretch into segments, one per thing they talked about, at most 4, in order. Skip stretches that are only gameplay noise, mumbling or garbled text; an empty list is a good answer when nothing was said. (3) topic: a short concrete name for the subject, the way a headline names it (\"Supermarket fridges hacked in France\", \"League ranked matchmaking\", \"Moving apartments\"), never vague (\"news\", \"gaming\", \"chatting\", \"the stream\") and never one of the kinds below (\"chat\", \"games\", \"life\"…): when nothing more specific fits, name what they were doing (\"Theme park rides\", \"Ranked teammates\"). When it is the same subject as one of the known topics, return that name exactly. (4) kind: news, games, tech, life, music, sports, money, internet, chat (talking with chat about the stream itself) or other. (5) title: a punchy headline for the segment, at most 70 characters. what: what they talked about, plainly, one or two sentences. take: their opinion or stance in one sentence, or null when they had none. quip: one line from the desk about the take or the situation, dry and a little savage, the way a late-night host would rib it, at most 160 characters. score: how much a reader would enjoy it, 1 (filler) to 5 (the best thing said all night). (6) quotes: 0 to 2 lines of the streamer's own words, copied character for character from the transcript lines (part of one line is fine), the funniest or most telling, each with its t. A quote is something they said themselves: never a sentence they read out from an article, a post, a tweet, a chat message or a video, even when it is in quotation marks in the source; when unsure, leave it out. When a quote is not in English, en is an English translation; otherwise en is null. (7) start and end are seconds from the lines. Write topic, title, what, take and quip in English whatever language they speak. In what and take, call the streamer by the name given in Streamer, never \"the streamer\" or \"Streamer\". (8) Never invent anything that is not in what you are given: no events, numbers, motives, relationships or words. Never accuse anyone of a crime or of wronging someone; a streamer naming another streamer is them talking, not a fight. Talk about what they said and did, never who they are: no jokes about looks, body, clothing, age, race, religion, sexuality, gender or disability, no sexual content about anyone, no threats, no self-harm jokes. Swearing is fine. Write every name exactly as given, same spelling and capitals, and call people by their name or \"they\": never guess a pronoun (no he, him, his, she, her). Privacy: never write or quote where anyone is, lives or will be (a named place, venue, event, school, town, address or time zone), their school or workplace, family members' names or when they will be somewhere; say \"a theme park\" or \"later\" instead. If someone seems to be under 18 (school, classmates, parents' rules), the desk is kind: no roasting them, nothing about school, and no quotes about where they are. Output only the JSON.",
          user_prompt: "Streamer: {{streamer}}\nStream title: {{title}}\nCategory: {{category}}\nKnown topics (reuse the exact name for the same subject):\n{{topics}}\nEarlier in this stream: {{earlier}}\n\nTranscript lines ([seconds into the stream] text):\n{{lines}}",
          default_route: 'live.creative', owner: 'live', metadata: { json_name: 'rundown_segment' },
      },
ddiff --git a/test/live-templates.test.js b/test/live-templates.test.js
iindex 229bde9..868c30c 100644
---- a/test/live-templates.test.js
++++ b/test/live-templates.test.js
@@@ -167,6 +167,7 @@ t.test('The Rundown: the segment prompt lists the lines and known topics; its ou
      assert.match(seg.system_prompt, /never "the streamer" or "Streamer"/);
      assert.match(seg.system_prompt, /theatrical one-liners in another/, 'a game character\'s English lines in a Greek stream are not the streamer');
      assert.match(seg.system_prompt, /commentators/);
++    assert.match(seg.system_prompt, /never one of the kinds below/);
      assert.match(seg.system_prompt, /never invent a meaning for them/);
      for (const k of ['live.rundown.segment', 'live.rundown.episode', 'live.rundown.topic']) {
          assert.match(tp(k).system_prompt, /Privacy: never write or quote where anyone is, lives or will be/, `${k}: no places, schools or time zones`);